Economy Artificial intelligence

AI in 2031: powerful, ordinary, and still our responsibility

Narayanan and Kapoor ask us to view AI as "normal technology". I largely agree, but "normal" should not sound reassuring: in three theses I explain why accountability becomes more urgent and where I expect things to stand in September 2031.

by Elia 24 September 2026 4 min read

Arvind Narayanan and Sayash Kapoor ask us to view AI as "normal technology": potentially transformative, but developed and deployed by people within existing organizations and institutions. Their essay challenges the idea that a more capable model automatically becomes an autonomous force that reshapes society overnight. I largely agree. The distinction between inventing a method, building a useful application, and getting people to trust and use that application is persuasive. Yet "normal" should not sound reassuring. Ordinary technologies can cause serious harm, and their benefits depend on who controls them. My view rests on three theses.

Thesis 1: Better models will change what is possible faster than organizations can change what they do. A model may perform well on a test and still fail at a job that involves messy data, unclear goals, and responsibility for errors. Narayanan and Kapoor's separation of AI methods, applications, and adoption captures this gap. In a US survey, Bick, Blandin, and Deming found rapid use of generative AI by late 2024, but estimated that it assisted only 1 to 5 percent of work hours. The distinction between trying a tool and restructuring work around it matters. A different kind of AI illustrates the stakes in a high consequence setting: an independent study of a sepsis prediction system found substantially worse performance in one hospital than its developer had reported. That result does not prove that generative AI will fail in hospitals. It shows why deployment needs local testing, monitoring, and a way to respond when a system is wrong. I therefore expect uneven progress: fast improvement in visible demonstrations, slower change where errors are costly.

Thesis 2: Calling AI normal makes accountability more urgent, not less. The strongest part of the normal technology argument is that it returns attention to decisions people can actually make: which systems to buy, which tasks to automate, who checks their output, and who bears the cost of mistakes. A controlled experiment by Noy and Zhang found that access to ChatGPT helped participants complete specific writing tasks 40 percent faster, with output rated 18 percent higher in quality. Those are meaningful gains, but an experiment on writing tasks cannot establish that every profession will gain equally or that workers will share the gains. Employers could use the same tool to help staff, intensify workloads, or cut entry level roles. I would require evidence from real use before delegating consequential decisions, and I would keep clear routes for human review and appeal. Narayanan and Kapoor are right to emphasize resilience and institutions; I am less confident than they are that ordinary incentives will reliably produce enough caution without external scrutiny.

Thesis 3: By September 2031, AI will be routine in many workflows, but independent authority will remain limited and uneven. My five year forecast is that drafting, translation, coding assistance, search, and document analysis will often begin with AI. Some software agents will complete bounded tasks when their actions can be checked or reversed. In medicine, law, and public administration, however, I expect the hardest part to remain proving that a system works in the particular setting where it is used. People will still sign off on many consequential actions, and organizations will spend substantial effort on evaluation, data access, training, and oversight. This is a prediction about deployment, not a claim that model research will slow down. A major breakthrough in reliability could make adoption faster than I expect, while poor incentives or weak institutions could make harms worse even without such a breakthrough. To judge this forecast in 2031, I would look beyond benchmark scores: What share of actual work is assisted or automated? Which decisions can be challenged? Who is accountable when an AI system fails?

The normal technology perspective gives me a useful way to hold two ideas together: AI can be powerful, and society still has choices about its use. The next five years will test whether we build institutions capable of making those choices well.